Images

Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F18/00—Pattern recognition
  • G06F18/20—Analysing
  • G06F18/24—Classification techniques
• • G06K9/00006—
• • G06K9/00228—
• • G06K9/6267—
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
  • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
  • G06V40/12—Fingerprints or palmprints
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
  • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
  • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
  • G06V40/161—Detection; Localisation; Normalisation
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
  • G06V40/70—Multimodal biometrics, e.g. combining information from different biometric modalities
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/20—Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/28—Constructional details of speech recognition systems
  • G10L15/285—Memory allocation or algorithm optimisation to reduce hardware requirements
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04M—TELEPHONIC COMMUNICATION
  • H04M1/00—Substation equipment, e.g. for use by subscribers
  • H04M1/72—Mobile telephones; Cordless telephones, i.e. devices for establishing wireless links to base stations without route selection
  • H04M1/724—User interfaces specially adapted for cordless or mobile telephones
  • H04M1/72448—User interfaces specially adapted for cordless or mobile telephones with means for adapting the functionality of the device according to specific conditions
  • H04M1/72454—User interfaces specially adapted for cordless or mobile telephones with means for adapting the functionality of the device according to specific conditions according to context-related or environment-related conditions
• • H04M1/72569—
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
  • G10L2015/226—Procedures used during a speech recognition process, e.g. man-machine dialogue using non-speech characteristics
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04M—TELEPHONIC COMMUNICATION
  • H04M2250/00—Details of telephonic subscriber devices
  • H04M2250/12—Details of telephonic subscriber devices including a sensor for measuring a physical value, e.g. temperature or motion

Definitions

• systems and methods according to the present technology use a smartphone to sense audio and/or visual information, and provide same to a first classifier module.
• the first classifier module characterizes the input audio-visual stimuli by type (e.g., music, speech, silence, video imagery, natural scene, face, etc.).
• a second classifier module processes other context information (which may include the output from the first classifier module), such as of day, day of week, location, calendar data, clock alarm status, motion sensors, Facebook status, etc., and outputs data characterizing a device state type, or scenario.
• a control rule module then issues control signals to one or more content recognition modules in accordance with the outputs from the two classifier modules.
• control signals can simply enable or disable the different recognition modules. Additionally, if a recognition module is enabled, the control signals can establish the frequency, or schedule, or other parameter(s), by which the module performs its recognition functions.
• Such arrangements conserve battery power, by not attempting operations that are unneeded or inappropriate to the context. Moreover, they aid other smartphone operations, since processing resources are not diverted to the idle recognition operations.
• FIG. 1 shows an illustrative embodiment that incorporates certain aspects of the present technology.
• FIG. 2 shows a few of the content recognition modules that may be used in the FIG. 1 embodiment.
• FIG. 3 is a block diagram of a process employing aspects of the present technology.
• FIG. 4 is a block diagram of an apparatus employing aspects of the present technology.
• FIG. 5 is an event controller table illustrating, for one embodiment, how different audio recognition agents are activated based on audio classification data.
• FIG. 6 is a flow chart illustrating, for one embodiment, how different audio recognition agents are activated based on audio classification data.
• FIG. 7 is an event controller table illustrating, for one embodiment, how different image recognition agents are activated based on outputs from light and motion sensors, and image classification data.
• FIG. 8 is a flow chart illustrating, for one embodiment, how different image recognition agents are activated based on outputs from light and motion sensors, and image classification data.
• an illustrative embodiment 10 that incorporates certain aspects of the present technology includes one or more microphones 12 , cameras 14 , audio-visual classifier modules 16 , second classifier modules 18 , control rule modules 20 , and content recognition modules 22 . These components may all be included in a smartphone. Alternatively, they may be distributed between different locations and/or different devices (including the cloud).
• One suitable smartphone is the Apple iPhone 4 device, which includes two cameras (one front-facing and one rear-facing), and two microphones.
• the audio-visual classifier module(s) 16 processes the data captured by the microphone(s) and/or camera(s), and classifies such audio-visual content by type.
• classification is the problem of identifying to which of a set of categories (sub-populations) a new observation belongs.
• the individual observations may be analyzed into a set of quantifiable properties, known variously as variables, features, etc. These properties may be categorical (e.g. “A”, “B”, “AB” or “O”, for blood type), ordinal (e.g. “large”, “medium” or “small”), etc.
• a familiar (although sometimes difficult) classification problem is identifying email as spam or not-spam.
• An algorithm or procedure that implements classification is known as a classifier.
• classification is performed based on a training set of data comprising observations (or instances) whose category membership is known.
• Classification in this sense, is regarded as an instance of supervised machine learning, i.e., learning where a training set of correctly-identified observations is available.
• the corresponding unsupervised procedure is known as clustering (or cluster analysis), and involves grouping data into categories based on some measure of inherent similarity (e.g. the distance between instances, considered as vectors in a multi-dimensional vector space).
• classification is regarded as including clustering.
• One illustrative classifier module 16 is an audio classifier, which categorizes input stimulus as speech, music, background/indeterminate, or silence. For the first three, the module also classifies the volume of the audio, as loud, mid-level, or quiet.
• a simple embodiment activates different content recognition modules, in accordance with the output of the audio classifier, as follows:
• the audio classifier module classifies the sensed audio as “silent” or “quiet background,” all three of the detailed content recognition modules are controlled to “off.” If the sensed audio is classified as music, the system activates the Nielsen audio watermark detector, and the Gracenote fingerprint engine, but leaves the Nuance speech recognition engine off.
• the audio watermark detector is activated, as is the speech recognition engine, but no fingerprint calculations are performed.
• the audio classifier identifies a loud or mid-level acoustic background, but is unable to further classify its type, the audio watermark detector, the fingerprint engine, and the speech recognition engine are all activated.
• Second classifier module 18 outputs device state type data, or scenario identification data, in accordance with context information.
• This context information may include the classification of the audio and/or visual environment (i.e., by the audio-visual classifier module(s) 16 , as shown by the dashed line in FIG. 1 ), and typically includes other information.
• This other context information can include, but is not limited to, time of day, day of week, location, calendar data, clock alarm status, motion and orientation sensor data, social network information (e.g., from Facebook), etc.
• Table II which expands the Table I information to include certain device state types as determined by the second classifier module (i.e., “away from office after work hours” and “at office during work hours”):
• Table II The first five rows of Table II are identical to Table I. They detail how the different modules are controlled when the user is away from the office, after work hours, given the noted audio environments.
• the final set of rows is different. These correspond to the device state type of being at the user's office, during work hours. As can be seen, only the speech recognition engine is ever activated in this context (i.e., when speech or loud background audio is sensed); the other modules are left idle regardless of the audio environment.
• the second classifier module 18 uses inputs such as time of day data and GPS data, in connection with reference data.
• This reference data establishes—for the particular smartphone user—the times of day that should be classified as work hours (e.g., 8 am-5 pm, Monday-Friday), and the location that should be classified as the office location (e.g., 45.4518° lat, ⁇ 122.7932° long, +/ ⁇ 0.0012 degrees).
• Table IIIA shows a more detailed scenario, as classified by the second classifier module 18 :
• Scenario 1 This confluence of circumstances is classed by the second classifier module as “Scenario 1.” It corresponds to the scenario in which the user is probably asleep (it is before 6:30 am on a weekday, and the alarm is set for 6:30; the smartphone is stationary in a quiet, dark environment).
• the control rules 20 associated with Scenario 1 cause all of the content recognition modules to be inactive.
• Scenario 2 corresponds to the user after waking up and before leaving home.
• the rules include instructions appropriate to this interval—while the user may be watching a morning news program on television, listening to the radio, or talking with a spouse.
• the Nielsen watermark detector is active, to allow the user to link to additional web content about something discussed on the television.
• the fingerprint engine is also active, so that the user might identify an appealing song that airs on the radio. Speech recognition may also be enabled, so that the spouse's verbal instructions to pick up ketchup, grapes, tin foil and postage stamps on the way home are transcribed for later reference.
• the user's smartphone also includes various visual content recognition capabilities, including facial recognition.
• the control rules specify that, in Scenario 2, facial recognition is disabled—as the user isn't expected to need prompting to recall faces of anyone encountered at home this early.
• Scenario 3 corresponds to the user's drive to work. No television audio is expected in this environment, so the Nielsen watermark detector is disabled. Song recognition and transcription of news from talk radio, however, may be helpful, so the fingerprint and speech recognition engines are enabled. Again, facial recognition is disabled.
• a different user may take the bus to work, instead of drive a car.
• Second scenario control rules for this user may be different. Without a car radio, song recognition is unneeded, so the fingerprint engine is disabled. However, the user sometimes overhears amusing conversations on the bus, so speech recognition is enabled so that any humorous dialog may be shared with work-mates. Occasionally, the user sees someone she should recognize on the bus—a parent of a child's soccer teammate, for example—but be unable to recall the name in this different environment. To prepare for this eventuality, the smartphone's facial recognition capability is loaded into memory and ready for operation, but does not process a frame of camera imagery until signaled by the user. (The signal may comprise the user holding the phone in a predetermined pose and saying the word “who.”)
• the smartphone is apparently lying face-up on a surface in the quiet work environment.
• the corresponding control rules specify that all recognition modules are disabled.
• the audio classifier indicates a change in audio environment—to mid-level or loud background sound, or speech
• the rules cause the phone to enable the speech recognition module. This provides the user with a transcribed record of any request or information she is given, or any instruction she issues, so that it may be referred-to later.
• Speech recognition can raise privacy concerns in some situations, including a work setting. Accordingly, the control rules cause the speech recognition module to issue an audible “beep” every thirty seconds when activated at work, to alert others that a recording is being made. In contrast, no “beep” alert is issued in the previously-discussed scenarios, because no recording of private third party speech is normally expected at home or in the car, and there is likely no expectation of privacy for speech overheard in a bus.
• Another datum of context that is processed by the illustrated second classifier module 18 is the number and identity of people nearby. “Nearby” may be within the range of a Bluetooth signal issued by a person's cell phone—typically 30 feet or less. Relative distance within this range can be assessed by strength of the Bluetooth signal, with a strong signal indicating, e.g., location within ten feet or less (i.e., “close”). Identity can be discerned—at least for familiar people—by reference to Bluetooth IDs for their known devices. Bluetooth IDs for devices owned by the user, family members, workmates, and other acquaintances, may be stored with the control rules to help discriminate well known persons from others.
• the rules can provide that speech recognition—if enabled—is performed without alert beeps, if the user is apparently solitary (i.e., no strong Bluetooth signals sensed, or only transitory signals—such as from strangers in nearby vehicles), or if the user is in the presence only of family members. However, if an unfamiliar strong Bluetooth signal is sensed when speech recognition is enabled, the system can instruct issuance of periodic alert beeps.
• the user's phone can present a user interface screen allowing the user to store this previously-unrecognized Bluetooth identifier.
• This UI allows the user to specify the identifier as corresponding to a family member, or to associate more particular identification information with the identifier (e.g., name and/or relationship). By such arrangement, beeping when not warranted is readily curtailed, and avoided when such circumstance recurs in the future.
• Scenario 5 a work meeting—is contextually similar to Scenario 4, except the audio classifier reports mid-level background audio, and the phone's location is in a meeting room.
• the speech recognition module is enabled, but corporate data retention policies require that transcripts of meetings be maintained only on corporate servers, so that they can be deleted after a retention period (e.g., 12 months) has elapsed.
• the control rules module 20 complies with this corporate policy, and immediately transmits the transcribed speech data to a corporate transcriptions database for storage—keeping no copy. Alert beeps are issued as a courtesy reminder of the recording.
• the rules cause the phone to beep only once every five minutes, instead of once every 30 seconds, to reduce the beeps' intrusiveness.
• the volume of the beeps can be reduced, based on a degree of social relationship between the user and the other sensed individual(s)—so that the beeps are louder when recording someone who is only distantly or not at all socially related to the user.
• Rules for face recognition in scenario 5 can vary depending on whether people sensed to be close-by are recognized by the user's phone. If all are recognized, the facial recognition module is not activated. However, if one or more close-by persons are not in the user's just-noted “friends” list (or within some more distant degree of relationship in a social network), then facial recognition is enabled as before—in an on-demand (rather than free-running) mode. (Alternatively, a different arrangement can be employed, e.g., with facial recognition activated if one or persons who have a certain type of social network linkage to the user, or absence thereof, are sensed to be present.)
• Scenario 6 finds the user in a subway, during the noon hour.
• the rules may be like those noted above for the bus commute.
• radio reception underground is poor.
• any facial recognition operation consults only facial eigenface reference data stored on the phone—rather than consulting the user's larger collection of Facebook or Picasa facial data, which are stored on cloud servers.
• Scenario 7 corresponds to a Friday evening birthday party. Lots of unfamiliar people are present, so the rules launch the facial recognition module in free-running mode—providing the user with the names of every recognized face within the field-of-view of any non-dark camera.
• This module relies on the user's Facebook and Picasa facial reference data stored in the cloud, as well as such data maintained in Facebook accounts of the user's Facebook friends. Speech recognition is disabled. Audio fingerprinting is enabled and—due to the party context—the phone has downloaded reference fingerprints for all the songs on Billboard's primary song lists (Hot 100, Billboard 200, and Hot 100 Airplay). Having this reference data cached on the phone allows much quicker operation of the song recognition application—at least for these 200+ songs.
• Fingerprint computation, watermark detection, and speech/facial recognition are computationally relatively expensive (“computationally heavy”). So are many classification tasks (e.g., speech/music classification). It is desirable to prevent such processes from running at a 100% duty cycle.
• One approach is to let the user decide when to run one or more heavy modules—with help from the output of one or more computationally light detectors. Adding additional steps to assess the signal quality prior to running one or more heavy detectors is another approach.
• a simple classifier e.g., a quietness classifier
• a simple classifier e.g., a quietness classifier
• Such module may indicate that there is a sudden change in the environment from a quiet state.
• Rules may call for activation of one or more heavy classifiers to determine whether the new audio environment is music or speech.
• the system may present a display screen with a “Confirm to Proceed” button that the user taps to undertake the classification. (There can also be an “Ignore” button.
• the system can have a default behavior, e.g., “Ignore,” if the user makes no selection within a pre-defined interval, such as ten seconds.)
• the user response to such prompts can be logged, and associated with different context information (including the sensitivity of the quietness classifier). Over time, this stored history data can be used to predict the circumstances in which the user instructs the heavy classifier to proceed. Action can then be taken based on such historical precedent, rather than always resorting to a user tap.
• the system can be self-learning, based on user interaction. For example, when a quietness classifier detects a change in loudness of amount “A,” it asks for the user's permission to enable a heavier classifier (for example, music versus speech classifier) or detector (e.g., watermark detector). If the user agrees, then this “A” level of loudness change is evidently at least sometimes of interest to the user. However, if over time, it becomes evident that the user uniformly refuses to activate the heavy classifier when the loudness changes by amount “A,” then the classifier can reset its threshold accordingly, and not ask the user for permission to activate the heavy module unless the loudness increases by “B” (where B>A). The quietness classifier thus learns to be less sensitive.
• a heavier classifier for example, music versus speech classifier
• detector e.g., watermark detector
• the quietness classifier learns to be more sensitive.
• FIG. 3 shows an arrangement using the foregoing principles.
• a microphone provides an ambient audio signal to a simple classifier, which produces an output based on a gross classification (e.g., silence or sound), based on a threshold audio level. If the classifier module switches from “silence” to “sound,” it causes the smartphone to present a user interface (UI) asking the user whether the system should invoke complex processing (e.g., speech recognition, speech/music classification, or other operation indicated by applicable rules). The system then acts in accordance with the user's instruction.
• UI user interface
• the current context is also stored in such history, as provided by a context classifier.
• the user history alone, may provide instructions as to how to respond in a given situation—without the need to ask the user.
• Another approach is to employ an additional classifier, to decide whether the current audio samples have a quality that merits further classification (i.e., with a heavy classifier). If the quality is judged to be insufficient, the heavy classifier is not activated (or is de-activated).
• Information-bearing signals such as speech and music—are commonly characterized by temporal variation, at least in spectral frequency content, and also generally in amplitude, when analyzed over brief time windows (e.g., 0.5 to 3 seconds).
• An additional classifier can listen for audio signals that are relatively uniform in spectral frequency content, and/or that are relatively uniform in average amplitude, over such a window interval.
• the signal may be regarded as interfering noise that unacceptably impairs the signal-to-noise ratio of the desired audio.
• the system discontinues heavy module processing until the interfering signal ceases.
• the just-discussed classifier detects the interval during which the jack-hammer is operated, and interrupts heavy audio processing during such period.
• Such classifier may similarly trigger when a loud train passes, or an air compressor operates, or even when a telephone rings—causing the system to change from its normal operation in these circumstances.
• Lu et al Another simple classifier relies on principles noted in Lu et al, SpeakerSense: Energy Efficient Unobtrusive Speaker Identification on Mobile Phones, Pervasive Computing Conf., 2011. Lu et al use a combination of signal energy (RMS) and zero crossing rate (ZCR) to distinguish human speech from other audio. While Lu et al use these parameters to identify speech, they can also be used to identify information-bearing signals more generally. (Or, stated otherwise, to flag audio passages that are likely lacking information, so that heavy processing modules can be disabled.)
• RMS signal energy
• ZCR zero crossing rate
• the additional classifier works after the detection of “sound change,” audio samples prior to the “sound change” can be used as an approximation of the background noise, and the audio sample after the “sound change” can be used as the background noise plus the useful signal. This gives a crude signal-to-noise ratio.
• the additional classifier can keep heavy modules in an idle state until this ratio exceeds a threshold value (e.g., 10 dB).
• Still another additional classifier to indicate the likely absence of an information-bearing signal—simply looks at a ratio of frequency components.
• the presence of high frequency signal components above a threshold amplitude is an indication of audio information.
• a ratio of energy in high frequency components e.g., above 2 KHz
• energy in low frequency components e.g., below 500 Hz
• One or more microphones provides a sensed audio signal to an audio screening classifier 30 (i.e., the “additional” classifier of the foregoing discussion).
• the microphone audio is optionally provided to a speech/music audio classifier 16 (as in FIG. 1 ) and to several heavy audio detector modules (e.g., watermark detector, speech recognition, etc.).
• the output of the audio screening classifier provides enable/disable control signals to the different heavy detectors.
• the audio screening classifier 30 provides the same control signal to all the heavy detectors, but in practical implementation, different controls signals may be generated for different detectors.
• the control signals from the audio screening classifier serve to disable the heavy detector(s), based on the audio sensed by the microphone.
• a context classifier 18 which operates like the second classifier module of FIG. 1 . It outputs signals indicating different context scenarios. These output data are provided to a control rules module 20 , which controls the mode of operation of the different heavy detector based on the identified scenario.
• FIG. 4 arrangements shows control of heavy detector modules, heavy classifier modules can be controlled by the same type of arrangement.
• Visual image classifiers e.g., facial recognition systems
• imagery having significant spatial variation in luminance (contrast/intensity) and/or hue (color/chrominance). If frames of imagery appear that are lacking in such variations, any heavy image processing module that would otherwise be operating should suspend its operation.
• a classifier can look for a series of image frames characterized by luminance or hue variation below a threshold, and interrupt heavy visual processing when such scene is detected.
• heavy visual processes are suspended when the user points a camera to a blank wall, or to the floor.
• uch action may also be taken based on smartphone orientation, e.g., with facial recognition only operative when the smartphone is oriented with its camera axis within 20 degrees of horizontal.
• Other threshold values can, of course, be used.
• a simple classifier can examine frame focus (e.g., by known metrics, such as high frequency content and contrast measures, or by a camera shake metric—provided by the phone's motion sensors), and disable facial recognition if the frame is likely blurred.
• Facial recognition can also be disabled if the subject is too distant to likely allow a correct identification. Thus, for example, if the phone's autofocus system indicates a focal distance of ten meters or more, facial recognition needn't be engaged.
• One technique relies on the smartphone's calendar app.
• the user's calendar, and phone clock indicate the user is at a meeting
• other participants in the user's proximity can be identified from the meeting participant data in the calendar app.
• location data which is short range-broadcast from the phone (or published from the phone to a common site), and used to indicate co-location with other phones.
• the location data can be derived from known techniques, including GPS, WiFi node identification, etc.
• a related approach relies on acoustic emitters that introduce subtle or inaudible background audio signals into an environment, which can be indicative of location.
• Software in a microphone-equipped device e.g., a smartphone app
• can listen for such signal e.g., above or below a range of human hearing, such as above 15-20 KHz
• the published information can include information conveyed by the sensed signal, e.g., identifying the emitting device or its owner, the device location and/or other context, etc.).
• the published information can also include information associated with the receiving device (e.g., identifying the device or its owner, the device location and/or other context, etc.). This allows a group of phones near each emitter to be identified. (Related technology is employed by the Shopkick service, and is detailed in patent publication US20110029370.)
• Bluetooth is presently preferred because—in addition to identifying nearby people, it also provides a communication channel with nearby phones. This enables the phones to collaborate in various tasks, including speech recognition, music fingerprinting, facial recognition, etc. For example, plural phones can exchange information about their respective battery states and/or other on-going processing tasks. An algorithm is then employed to select one phone to perform a particular task (e.g., the one with the most battery life remaining is selected to perform watermark decoding or facial recognition). This phone then transmits the results of its task—or related information based thereon—to the other phones (by Bluetooth or otherwise).
• 3D image modeling based on camera data from two or more different phones, each with a different view of a subject.
• a particular application is facial recognition, where two or more different views of a person allow a 3D facial model to be generated. Facial recognition can then be based on the 3D model information—yielding a more certain identification than 2D facial recognition affords.
• a phone processes ambient audio/visual stimulus in connection with phone-specific information, allowing different phones to provide different results.
• phone-specific information includes history, contacts, computing context, user context, physical context, etc. See, e.g., published application 20110161076 and copending application Ser. No. 13/174,258, filed Jun. 30, 2011 (now U.S. Pat. No. 8,831,279).
• image processing different phones may have better or worse views of the subject.
• collaborating phones can send the audio/imagery they captured to one or more other phones for processing.
• a phone having Facebook access to useful facial recognition data may not be the phone with the best view of a person to be identified. If plural phones each captures data, and shares such data (or information based thereon, e.g., eigenface data) to the other phones, results may be achieved that are better than any phone—by itself—could manage.
• devices may communicate other than by Bluetooth.
• NFC and WiFi are two such alternatives.
• Bluetooth was also noted as a technique for determining that a user is in a vehicle. Again, other arrangements can be employed.
• GPS can establish that the user is following established roads, and is moving at speeds above that associated with walking or biking.
• terrain elevation can be considered. If the terrain is generally flat, or if the traveler is going uphill, a sustained speed of more than 20 mph may distinguish motorized transport from bicycling. However, if the user is following a road down a steep downhill incline, then a sustained speed of more than 35 mph may be used to establish motorized travel with certainty.
• two or more phones report, e.g., by a shared short-range context broadcast, that they are each following the same geo-location-track, at the same speed, then the users of the two phones can conclude that they are traveling on the same conveyance—whether car, bus, bike, etc.
• the cloud can serve as a recipient for such information reported by the smartphones, can make determinations, e.g., about correlation between devices.
• the information shared may be of such a character (e.g., acceleration, captured audio) that privacy concerns do not arise—given the short range of transmission involved.
• OCR optical character recognition
• barcode decoding
• a user-facing camera can be used to assess emotions of a user (or user response to information presented on the smartphone screen), and tailor operation of the phone—including use of the other camera—accordingly.
• a user-facing camera can also detect the user's eye position. Operation of the phone can thereby be controlled. For example, instead of switching between “portrait” and “landscape” display modes based on the phone's position sensors, this screen display mode can be controlled based on the orientation of the user's eyes.
• this screen display mode can be controlled based on the orientation of the user's eyes.
• the phone can operate its display in the “portrait” mode. If the user turns the phone ninety degrees (i.e., so that its long axis is parallel to the axis between the users' eyes), the phone switches its display mode to “landscape.”
• the screen mode switches to follow the relative orientation of the axis between the user's eyes, relative to the screen axis. (That is, if the long axis of the phone is parallel with the axis between the user's eyes, landscape mode is used; and vice versa.)
• the phone is equipped with stereo cameras (i.e., two cameras with fields of view that overlap)
• the two views can be used for distance determination to any point in the frame (i.e., range finding).
• the distance information can be used by the phone processor to guide the user to move the phone closer to, or further from, the intended subject in order to achieve best results.
• a phone may seek to identify an audio scene by reference to sensed audio.
• a meeting room scene may be acoustically characterized by a quiet background, with distinguishable human speech, and with occasional sound source transition (different people speaking alternatively).
• a home scene with the user and her husband may be acoustically characterized by a mid-level background audio (perhaps music or television), and by two different voices speaking alternately.
• a crowded convention center may be characterized by a high-level background sound, with many indistinguishable human voices, and occasionally the user's voice or another.
• two or more smartphones can act and cooperate in different ways. For example, if the scene has been identified as a meeting, the user's phone can automatically check-in for the room, indicating that the meeting room is occupied. (Calendar programs are often used for this, but impromptu meetings may occupy rooms without advance scheduling. The smartphone can enter the meeting on the calendar—booking the room against competing reservations—after the meeting has started.)
• the phone may communicate with a laptop or other device controlling a Powerpoint slide presentation, to learn the number of slides in the deck being reviewed, and the slide currently being displayed.
• the laptop or the phone can compute how quickly the slides are being advanced, and extrapolate when the meeting will conclude. (E.g., if a deck has 30 slides, and it has taken 20 minutes to get through 15 slides, a processor can figure it will take a further 20 minutes to get through the final 15 slides. Adding 10 minutes at the end for a wrap-up discussion, it can figure that the meeting will conclude in 30 minutes.)
• This information can be shared with participants, or posted to the calendar app to indicate when the room might become available.
• the two phones can exchange domestic information (e.g., shopping list information, social calendar data, bills to be paid soon, etc.).
• domestic information e.g., shopping list information, social calendar data, bills to be paid soon, etc.
• a phone can initiate automatic electronic business card exchange (e.g., v-card), if it senses the user having a chat with another person, and the phone does not already have contact information for the other person (e.g., as identified by Bluetooth-indicated cell phone number, or otherwise).
• automatic electronic business card exchange e.g., v-card
• the user's phone might also check the public calendars of people with whom the user talks, to identify those with similar transportation requirements (e.g., a person whose flight departs from the same airport as the user's departing flight, with a flight time within 30 minutes of the user's flight). Such information can then be brought to the attention of the user, e.g., with an audible or tactile alert.
• Tasks can be referred to the cloud based on various factors.
• An example is use cloud processing for “easy to transmit” data (i.e., small size) and “hard to calculate” tasks (i.e., computationally complex). Cloud processing is often best suited for tasks that don't require extensive local knowledge (e.g., device history and other information stored on the device).
• a traveler flying to San Francisco for a conference who needs to commute to a conference center hotel downtown.
• the user's phone On landing at the airport, the user's phone sends the address of the downtown hotel/conference center to a cloud server.
• the cloud server has knowledge of real-time traffic information, construction delays, etc.
• the server calculates the optimal route under various constraints, e.g., shortest-time route, shortest-distance route, most cost-effective route, etc. If the user arrives at the airport only 20 minutes before the conference begins, the phone suggests taking a taxi (perhaps suggesting sharing a taxi with others that it senses have the same destination—perhaps others who also have a third-party trustworthiness score exceeding “good”).
• the smartphone can rely on audio and imagery collected by public sensors, such as surveillance cameras in a parking garage, mall, convention center, or a home security system. This information can be part of the “big computation” provided by cloud processing. Or the data can be processed exclusively by the smartphone, such as helping the user find where she parked her yellow Nissan Leaf automobile in a crowded parking lot.
• information from the user's social networking accounts can be used as input to the arrangements detailed herein (e.g., as context information).
• information from the accounts of people that the user encounters e.g., at work, home, conferences, etc.
• information output from the detailed arrangements can be posted automatically to the user's social networking account(s).
• facial recognition has a number of uses.
• One, noted above, is as a memory aid—prompting the user with a name of an acquaintance.
• the user's smartphone may broadcast certain private information only if it recognizes a nearby person as a friend (e.g., by reference to the user's list of Friends on Facebook).
• Facial recognition can also be used to tag images of a person with the person's name and other information.
• a user's smartphone broadcasts one or more high quality facial portraits of the user, or associated eigenface data.
• Another smartphone user can snap a poor picture of the user. That smartphone compares the snapped image with high quality image data (or eigenface data) received over Bluetooth from the user, and can confirm that the poor picture and the received image data correspond to the same individual.
• the other smartphone uses the received image data in lieu of the poor picture, e.g., for facial recognition, or to illustrate a Contacts list, or for any other purpose where the user's photo might be employed.
• FIG. 5 shows an event controller table for another audio embodiment, indicating how two digital watermark decoders (one tailored for watermarks commonly found in music, another tailored for watermarks commonly found in broadcast speech) are controlled, based on classifier data categorizing input audio as likely silence, speech, and/or music.
• FIG. 6 shows a corresponding flow chart.
• FIG. 7 shows an event controller table for another embodiment—this one involving imagery.
• This arrangement shows how different recognition modules (1D barcode, 2D barcode, image watermark, image fingerprint, and OCR) are controlled in accordance with different sensor information.
• Sensors can encompass logical sensors, such as classifiers.
• the system includes a light sensor, and a motion sensor. Additionally, one or more image classifiers outputs information identifying the imagery as likely depicting text, a 1D barcode, or a 2D barcode.
• the image watermark decoding module and the image fingerprint module, are activated based on certain combinations of outputs from the classifier(s) (e.g., when none or all of the three types of classified images is identified).
• FIG. 8 shows a corresponding flow chart.
• Audio classification problem is often termed as content-based classification/retrieval, or audio segmentation. There are two basic issues in this work: feature selection and classifier selection.
• Audio understanding can be based on features in three layers: low-level acoustic characteristics, intermediate-level audio signatures associated with different sounding objects, and high level semantic models of audio in different scene classes”; and “classification based on these low-level features alone may not be accurate, but the error can be addressed in a higher layer by examining the structure underlying a sequence of continuous audio clips.”
• Lu in [9] used low-level audio features, including 8 order MFCCs and several other perceptual features, as well as kernel SVM (support vector machine) as the classifier in a cascaded scheme.
• the work in [10] also included perceptual features and used different classifiers in a cascaded classification scheme, including k-NN, LSP VQ and rule based methods (for smoothing). In this paper they used dynamic feature sets (use different features) for classifying different classes.
• Arrangements particularly focused on speech/music discrimination include [19] and [20].
• Exemplary classifiers also include those detailed in patent publications 20020080286 (British Telecomm), 20020080286 (NEC), 20020080286 (Philips), 20030009325 (DeutscheInstitut), 20040210436 (Microsoft), 20100257129 and 20120109643 (Google), and U.S. Pat. No. 5,712,953 (Hewlett-Packard).
• head-worn devices e.g., Google Glass goggles
• other unobtrusive sensor platforms will eventually replace today's smartphones.
• the present technology can be used with such other forms of devices.
• smartphone should be construed to encompass all such devices, even those that are not strictly-speaking cellular, nor telephones.
• each includes one or more processors, one or more memories (e.g. RAM), storage (e.g., a disk or flash memory), a user interface (which may include, e.g., a keypad, a TFT LCD or OLED display screen, touch or other gesture sensors, a camera or other optical sensor, a compass sensor, a 3D magnetometer, a 3-axis accelerometer, a microphone, etc., together with software instructions for providing a graphical user interface), interconnections between these elements (e.g., buses), and an interface for communicating with other devices (which may be wireless, such as GSM, CDMA, W-CDMA, CDMA2000, TDMA, EV-DO, HSDPA, WiFi, WiMax, or Bluetooth, and/or wired, such as through an Ethernet local area network, a T-1 internet connection, etc.).
• memories e.g. RAM
• storage e.g., a disk or flash memory
• a user interface which may include, e.g., a key
• the processes and system components detailed in this specification may be implemented as instructions for computing devices, including general purpose processor instructions for a variety of programmable processors, including microprocessors, graphics processing units (GPUs, such as the nVidia Tegra APX 2600), digital signal processors (e.g., the Texas Instruments TMS320 series devices), etc. These instructions may be implemented as software, firmware, etc. These instructions can also be implemented to various forms of processor circuitry, including programmable logic devices, FPGAs (e.g., the noted Xilinx Virtex series devices), FPOAs (e.g., the noted PicoChip devices), and application specific circuits—including digital, analog and mixed analog/digital circuitry.
• FPGAs e.g., the noted Xilinx Virtex series devices
• FPOAs e.g., the noted PicoChip devices
• application specific circuits including digital, analog and mixed analog/digital circuitry.
• Execution of the instructions can be distributed among processors and/or made parallel across processors within a device or across a network of devices. Transformation of content signal data may also be distributed among different processor and memory devices. References to “processors” or “modules” should be understood to refer to functionality, rather than requiring a particular form of hardware and/or software implementation.
• Smartphones and other devices can include software modules for performing the different functions and acts.
• Known artificial intelligence systems and techniques can be employed to make the inferences, conclusions, and other determinations noted above.
• each device includes operating system software that provides interfaces to hardware resources and general purpose functions, and also includes application software which can be selectively invoked to perform particular tasks desired by a user.
• Known browser software, communications software, and media processing software can be adapted for many of the uses detailed herein.
• Software and hardware configuration data/instructions are commonly stored as instructions in one or more data structures conveyed by tangible media, such as magnetic or optical discs, memory cards, ROM, etc., which may be accessed across a network.
• Some embodiments may be implemented as embedded systems—a special purpose computer system in which the operating system software and the application software is indistinguishable to the user (e.g., as is commonly the case in basic cell phones).
• the functionality detailed in this specification can be implemented in operating system software, application software and/or as embedded system software.
• Bluetooth technology to indicate proximity to, and identity of, nearby persons is illustrative only. Many alternative technologies are known to perform one or both of these functions, and can be readily substituted.

Abstract

Description

Claims (20)

Priority Applications (5)

Applications Claiming Priority (15)

Related Parent Applications (2)

Related Child Applications (1)

Publications (2)

Family

ID=47914762

Family Applications (6)

Family Applications Before (1)

Family Applications After (4)

Country Status (6)

Cited By (2)

Families Citing this family (85)

Citations (65)

Family Cites Families (54)

• 2012
  • 2012-09-07 CN CN201280054460.1A patent/CN103918247B/en active Active
  • 2012-09-07 US US13/607,095 patent/US9196028B2/en active Active
  • 2012-09-07 EP EP12833294.7A patent/EP2758956B1/en active Active
  • 2012-09-07 WO PCT/US2012/054232 patent/WO2013043393A1/en unknown
  • 2012-09-07 KR KR1020147010125A patent/KR20140064969A/en not_active Application Discontinuation
  • 2012-09-07 JP JP2014531853A patent/JP6251906B2/en active Active
• 2015
  • 2015-11-20 US US14/947,008 patent/US9595258B2/en active Active
• 2017
  • 2017-03-01 US US15/446,837 patent/US20170243584A1/en not_active Abandoned
  • 2017-09-21 US US15/711,357 patent/US10199042B2/en active Active
• 2019
  • 2019-01-30 US US16/262,634 patent/US10510349B2/en active Active
  • 2019-12-10 US US16/709,463 patent/US10930289B2/en active Active

Patent Citations (67)

Non-Patent Citations (28)

Also Published As

Similar Documents

Legal Events